7 #include <movit/effect_chain.h>
8 #include <movit/effect_util.h>
9 #include <movit/flat_input.h>
10 #include <movit/image_format.h>
11 #include <movit/init.h>
12 #include <movit/resource_pool.h>
13 #include <movit/util.h>
17 #include <sys/resource.h>
20 #include <condition_variable>
31 #include "DeckLinkAPI.h"
33 #include "alsa_output.h"
34 #include "bmusb/bmusb.h"
35 #include "bmusb/fake_capture.h"
37 #include "decklink_capture.h"
39 #include "disk_space_estimator.h"
41 #include "input_mapping.h"
42 #include "pbo_frame_allocator.h"
43 #include "ref_counted_gl_sync.h"
44 #include "resampling_queue.h"
46 #include "video_encoder.h"
51 using namespace movit;
53 using namespace std::chrono;
54 using namespace std::placeholders;
55 using namespace bmusb;
57 Mixer *global_mixer = nullptr;
58 bool uses_mlock = false;
62 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
65 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) { // :-)
66 input_state->buffered_frames[card_index][frame_num] =
67 input_state->buffered_frames[card_index][frame_num - 1];
69 input_state->buffered_frames[card_index][0] = { frame, field_num };
71 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
72 input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
79 void QueueLengthPolicy::update_policy(int queue_length)
81 if (queue_length < 0) { // Starvation.
82 if (been_at_safe_point_since_last_starvation && safe_queue_length < 5) {
84 fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frames\n",
85 card_index, safe_queue_length);
87 frames_with_at_least_one = 0;
88 been_at_safe_point_since_last_starvation = false;
91 if (queue_length > 0) {
92 if (queue_length >= int(safe_queue_length)) {
93 been_at_safe_point_since_last_starvation = true;
95 if (++frames_with_at_least_one >= 1000 && safe_queue_length > 0) {
97 fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frames\n",
98 card_index, safe_queue_length);
99 frames_with_at_least_one = 0;
102 frames_with_at_least_one = 0;
106 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
108 num_cards(num_cards),
109 mixer_surface(create_surface(format)),
110 h264_encoder_surface(create_surface(format)),
111 audio_mixer(num_cards)
113 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
116 // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
117 // will be halved when sampling them, and we need to compensate here.
118 movit_texel_subpixel_precision /= 2.0;
120 resource_pool.reset(new ResourcePool);
121 theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
122 for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
123 output_channel[i].parent = this;
124 output_channel[i].channel = i;
127 ImageFormat inout_format;
128 inout_format.color_space = COLORSPACE_sRGB;
129 inout_format.gamma_curve = GAMMA_sRGB;
131 // Display chain; shows the live output produced by the main chain (its RGBA version).
132 display_chain.reset(new EffectChain(global_flags.width, global_flags.height, resource_pool.get()));
134 display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, global_flags.width, global_flags.height); // FIXME: GL_UNSIGNED_BYTE is really wrong.
135 display_chain->add_input(display_input);
136 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
137 display_chain->set_dither_bits(0); // Don't bother.
138 display_chain->finalize();
140 video_encoder.reset(new VideoEncoder(resource_pool.get(), h264_encoder_surface, global_flags.va_display, global_flags.width, global_flags.height, &httpd, global_disk_space_estimator));
142 // Start listening for clients only once VideoEncoder has written its header, if any.
145 // First try initializing the then PCI devices, then USB, then
146 // fill up with fake cards until we have the desired number of cards.
147 unsigned num_pci_devices = 0;
148 unsigned card_index = 0;
151 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
152 if (decklink_iterator != nullptr) {
153 for ( ; card_index < num_cards; ++card_index) {
155 if (decklink_iterator->Next(&decklink) != S_OK) {
159 configure_card(card_index, new DeckLinkCapture(decklink, card_index), /*is_fake_capture=*/false);
162 decklink_iterator->Release();
163 fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
165 fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
168 unsigned num_usb_devices = BMUSBCapture::num_cards();
169 for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
170 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
171 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
172 configure_card(card_index, capture, /*is_fake_capture=*/false);
174 fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
176 unsigned num_fake_cards = 0;
177 for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
178 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
179 configure_card(card_index, capture, /*is_fake_capture=*/true);
182 if (num_fake_cards > 0) {
183 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
186 BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
187 BMUSBCapture::start_bm_thread();
189 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
190 cards[card_index].queue_length_policy.reset(card_index);
193 // Set up stuff for NV12 conversion.
195 // Note: Due to the horizontally co-sited chroma/luma samples in H.264
196 // (chrome position is left for horizontal and center for vertical),
197 // we need to be a bit careful in our subsampling. A diagram will make
198 // this clearer, showing some luma and chroma samples:
213 // Clearly, the rightmost chroma sample here needs to be equivalent to
214 // b/4 + c/2 + d/4. (We could also implement more sophisticated filters,
215 // of course, but as long as the upsampling is not going to be equally
216 // sophisticated, it's probably not worth it.) If we sample once with
217 // no mipmapping, we get just c, ie., no actual filtering in the
218 // horizontal direction. (For the vertical direction, we can just
219 // sample in the middle to get the right filtering.) One could imagine
220 // we could use mipmapping (assuming we can create mipmaps cheaply),
221 // but then, what we'd get is this:
236 // which ends up sampling equally from a and b, which clearly isn't right. Instead,
237 // we need to do two (non-mipmapped) chroma samples, both hitting exactly in-between
240 // Sampling in-between b and c gives us the sample (b+c)/2, and similarly for c and d.
241 // Taking the average of these gives of (b+c)/4 + (c+d)/4 = b/4 + c/2 + d/4, which is
242 // exactly what we want.
244 // See also http://www.poynton.com/PDFs/Merging_RGB_and_422.pdf, pages 6–7.
247 string cbcr_vert_shader =
250 "in vec2 position; \n"
251 "in vec2 texcoord; \n"
252 "out vec2 tc0, tc1; \n"
253 "uniform vec2 foo_chroma_offset_0; \n"
254 "uniform vec2 foo_chroma_offset_1; \n"
258 " // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is: \n"
260 " // 2.000 0.000 0.000 -1.000 \n"
261 " // 0.000 2.000 0.000 -1.000 \n"
262 " // 0.000 0.000 -2.000 -1.000 \n"
263 " // 0.000 0.000 0.000 1.000 \n"
264 " gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0); \n"
265 " vec2 flipped_tc = texcoord; \n"
266 " tc0 = flipped_tc + foo_chroma_offset_0; \n"
267 " tc1 = flipped_tc + foo_chroma_offset_1; \n"
269 string cbcr_frag_shader =
271 "in vec2 tc0, tc1; \n"
272 "uniform sampler2D cbcr_tex; \n"
273 "out vec4 FragColor; \n"
275 " FragColor = 0.5 * (texture(cbcr_tex, tc0) + texture(cbcr_tex, tc1)); \n"
277 vector<string> frag_shader_outputs;
278 cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
285 cbcr_vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
286 cbcr_texture_sampler_uniform = glGetUniformLocation(cbcr_program_num, "cbcr_tex");
287 cbcr_position_attribute_index = glGetAttribLocation(cbcr_program_num, "position");
288 cbcr_texcoord_attribute_index = glGetAttribLocation(cbcr_program_num, "texcoord");
290 if (global_flags.enable_alsa_output) {
291 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
297 resource_pool->release_glsl_program(cbcr_program_num);
298 glDeleteBuffers(1, &cbcr_vbo);
299 BMUSBCapture::stop_bm_thread();
301 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
303 unique_lock<mutex> lock(bmusb_mutex);
304 cards[card_index].should_quit = true; // Unblock thread.
305 cards[card_index].new_frames_changed.notify_all();
307 cards[card_index].capture->stop_dequeue_thread();
310 video_encoder.reset(nullptr);
313 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, bool is_fake_capture)
315 printf("Configuring card %d...\n", card_index);
317 CaptureCard *card = &cards[card_index];
318 if (card->capture != nullptr) {
319 card->capture->stop_dequeue_thread();
320 delete card->capture;
322 card->capture = capture;
323 card->is_fake_capture = is_fake_capture;
324 card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
325 if (card->frame_allocator == nullptr) {
326 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, global_flags.width, global_flags.height)); // 8 MB.
328 card->capture->set_video_frame_allocator(card->frame_allocator.get());
329 if (card->surface == nullptr) {
330 card->surface = create_surface_with_same_format(mixer_surface);
332 while (!card->new_frames.empty()) card->new_frames.pop();
333 card->fractional_samples = 0;
334 card->last_timecode = -1;
335 card->capture->configure_card();
337 DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
338 audio_mixer.reset_resampler(device);
339 audio_mixer.set_display_name(device, card->capture->get_description());
340 audio_mixer.trigger_state_changed_callback();
346 int unwrap_timecode(uint16_t current_wrapped, int last)
348 uint16_t last_wrapped = last & 0xffff;
349 if (current_wrapped > last_wrapped) {
350 return (last & ~0xffff) | current_wrapped;
352 return 0x10000 + ((last & ~0xffff) | current_wrapped);
358 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
359 FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
360 FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
362 DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
363 CaptureCard *card = &cards[card_index];
365 if (is_mode_scanning[card_index]) {
366 if (video_format.has_signal) {
367 // Found a stable signal, so stop scanning.
368 is_mode_scanning[card_index] = false;
370 static constexpr double switch_time_s = 0.5; // Should be enough time for the signal to stabilize.
371 steady_clock::time_point now = steady_clock::now();
372 double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
373 if (sec_since_last_switch > switch_time_s) {
374 // It isn't this mode; try the next one.
375 mode_scanlist_index[card_index]++;
376 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
377 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
378 last_mode_scan_change[card_index] = now;
383 int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
384 assert(frame_length > 0);
386 size_t num_samples = (audio_frame.len > audio_offset) ? (audio_frame.len - audio_offset) / audio_format.num_channels / (audio_format.bits_per_sample / 8) : 0;
387 if (num_samples > OUTPUT_FREQUENCY / 10) {
388 printf("Card %d: Dropping frame with implausible audio length (len=%d, offset=%d) [timecode=0x%04x video_len=%d video_offset=%d video_format=%x)\n",
389 card_index, int(audio_frame.len), int(audio_offset),
390 timecode, int(video_frame.len), int(video_offset), video_format.id);
391 if (video_frame.owner) {
392 video_frame.owner->release_frame(video_frame);
394 if (audio_frame.owner) {
395 audio_frame.owner->release_frame(audio_frame);
400 int dropped_frames = 0;
401 if (card->last_timecode != -1) {
402 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
405 // Number of samples per frame if we need to insert silence.
406 // (Could be nonintegral, but resampling will save us then.)
407 const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
409 if (dropped_frames > MAX_FPS * 2) {
410 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
411 card_index, card->last_timecode, timecode);
412 audio_mixer.reset_resampler(device);
414 } else if (dropped_frames > 0) {
415 // Insert silence as needed.
416 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
417 card_index, dropped_frames, timecode);
421 success = audio_mixer.add_silence(device, silence_samples, dropped_frames, frame_length);
425 audio_mixer.add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, frame_length);
427 // Done with the audio, so release it.
428 if (audio_frame.owner) {
429 audio_frame.owner->release_frame(audio_frame);
432 card->last_timecode = timecode;
434 size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
435 if (video_frame.len - video_offset == 0 ||
436 video_frame.len - video_offset != expected_length) {
437 if (video_frame.len != 0) {
438 printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
439 card_index, video_frame.len - video_offset, expected_length);
441 if (video_frame.owner) {
442 video_frame.owner->release_frame(video_frame);
445 // Still send on the information that we _had_ a frame, even though it's corrupted,
446 // so that pts can go up accordingly.
448 unique_lock<mutex> lock(bmusb_mutex);
449 CaptureCard::NewFrame new_frame;
450 new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
451 new_frame.length = frame_length;
452 new_frame.interlaced = false;
453 new_frame.dropped_frames = dropped_frames;
454 card->new_frames.push(move(new_frame));
455 card->new_frames_changed.notify_all();
460 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
462 unsigned num_fields = video_format.interlaced ? 2 : 1;
463 steady_clock::time_point frame_upload_start;
464 if (video_format.interlaced) {
465 // Send the two fields along as separate frames; the other side will need to add
466 // a deinterlacer to actually get this right.
467 assert(video_format.height % 2 == 0);
468 video_format.height /= 2;
469 assert(frame_length % 2 == 0);
472 frame_upload_start = steady_clock::now();
474 userdata->last_interlaced = video_format.interlaced;
475 userdata->last_has_signal = video_format.has_signal;
476 userdata->last_is_connected = video_format.is_connected;
477 userdata->last_frame_rate_nom = video_format.frame_rate_nom;
478 userdata->last_frame_rate_den = video_format.frame_rate_den;
479 RefCountedFrame frame(video_frame);
481 // Upload the textures.
482 size_t cbcr_width = video_format.width / 2;
483 size_t cbcr_offset = video_offset / 2;
484 size_t y_offset = video_frame.size / 2 + video_offset / 2;
486 for (unsigned field = 0; field < num_fields; ++field) {
487 // Put the actual texture upload in a lambda that is executed in the main thread.
488 // It is entirely possible to do this in the same thread (and it might even be
489 // faster, depending on the GPU and driver), but it appears to be trickling
490 // driver bugs very easily.
492 // Note that this means we must hold on to the actual frame data in <userdata>
493 // until the upload command is run, but we hold on to <frame> much longer than that
494 // (in fact, all the way until we no longer use the texture in rendering).
495 auto upload_func = [field, video_format, y_offset, cbcr_offset, cbcr_width, userdata]() {
496 unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22);
498 if (userdata->tex_y[field] == 0 ||
499 userdata->tex_cbcr[field] == 0 ||
500 video_format.width != userdata->last_width[field] ||
501 video_format.height != userdata->last_height[field]) {
502 // We changed resolution since last use of this texture, so we need to create
503 // a new object. Note that this each card has its own PBOFrameAllocator,
504 // we don't need to worry about these flip-flopping between resolutions.
505 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
507 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
509 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
511 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
513 userdata->last_width[field] = video_format.width;
514 userdata->last_height[field] = video_format.height;
517 GLuint pbo = userdata->pbo;
519 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo);
522 size_t field_y_start = y_offset + video_format.width * field_start_line;
523 size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
525 if (global_flags.flush_pbos) {
526 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_y_start, video_format.width * video_format.height);
528 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_cbcr_start, cbcr_width * video_format.height * sizeof(uint16_t));
532 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
534 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_cbcr_start));
536 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
538 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_y_start));
540 glBindTexture(GL_TEXTURE_2D, 0);
542 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
547 // Don't upload the second field as fast as we can; wait until
548 // the field time has approximately passed. (Otherwise, we could
549 // get timing jitter against the other sources, and possibly also
550 // against the video display, although the latter is not as critical.)
551 // This requires our system clock to be reasonably close to the
552 // video clock, but that's not an unreasonable assumption.
553 steady_clock::time_point second_field_start = frame_upload_start +
554 nanoseconds(frame_length * 1000000000 / TIMEBASE);
555 this_thread::sleep_until(second_field_start);
559 unique_lock<mutex> lock(bmusb_mutex);
560 CaptureCard::NewFrame new_frame;
561 new_frame.frame = frame;
562 new_frame.length = frame_length;
563 new_frame.field = field;
564 new_frame.interlaced = video_format.interlaced;
565 new_frame.upload_func = upload_func;
566 new_frame.dropped_frames = dropped_frames;
567 new_frame.received_timestamp = video_frame.received_timestamp; // Ignore the audio timestamp.
568 card->new_frames.push(move(new_frame));
569 card->new_frames_changed.notify_all();
574 void Mixer::bm_hotplug_add(libusb_device *dev)
576 lock_guard<mutex> lock(hotplug_mutex);
577 hotplugged_cards.push_back(dev);
580 void Mixer::bm_hotplug_remove(unsigned card_index)
582 cards[card_index].new_frames_changed.notify_all();
585 void Mixer::thread_func()
587 eglBindAPI(EGL_OPENGL_API);
588 QOpenGLContext *context = create_context(mixer_surface);
589 if (!make_current(context, mixer_surface)) {
594 // Start the actual capture. (We don't want to do it before we're actually ready
595 // to process output frames.)
596 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
597 cards[card_index].capture->start_bm_capture();
600 steady_clock::time_point start, now;
601 start = steady_clock::now();
604 int stats_dropped_frames = 0;
606 while (!should_quit) {
607 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
608 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
609 int num_samples[MAX_VIDEO_CARDS] = { 0 };
611 unsigned master_card_index = theme->map_signal(master_clock_channel);
612 assert(master_card_index < num_cards);
614 get_one_frame_from_each_card(master_card_index, new_frames, has_new_frame, num_samples);
615 schedule_audio_resampling_tasks(new_frames[master_card_index].dropped_frames, num_samples[master_card_index], new_frames[master_card_index].length);
616 stats_dropped_frames += new_frames[master_card_index].dropped_frames;
618 handle_hotplugged_cards();
620 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
621 if (card_index == master_card_index || !has_new_frame[card_index]) {
624 if (new_frames[card_index].frame->len == 0) {
625 ++new_frames[card_index].dropped_frames;
627 if (new_frames[card_index].dropped_frames > 0) {
628 printf("Card %u dropped %d frames before this\n",
629 card_index, int(new_frames[card_index].dropped_frames));
633 // If the first card is reporting a corrupted or otherwise dropped frame,
634 // just increase the pts (skipping over this frame) and don't try to compute anything new.
635 if (new_frames[master_card_index].frame->len == 0) {
636 ++stats_dropped_frames;
637 pts_int += new_frames[master_card_index].length;
641 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
642 if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
645 CaptureCard::NewFrame *new_frame = &new_frames[card_index];
646 assert(new_frame->frame != nullptr);
647 insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
650 // The new texture might need uploading before use.
651 if (new_frame->upload_func) {
652 new_frame->upload_func();
653 new_frame->upload_func = nullptr;
657 int64_t frame_duration = new_frames[master_card_index].length;
658 render_one_frame(frame_duration);
660 pts_int += frame_duration;
662 now = steady_clock::now();
663 double elapsed = duration<double>(now - start).count();
664 if (frame % 100 == 0) {
665 printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
666 frame, stats_dropped_frames, elapsed, frame / elapsed,
667 1e3 * elapsed / frame);
668 // chain->print_phase_timing();
670 // Check our memory usage, to see if we are close to our mlockall()
671 // limit (if at all set).
673 if (getrusage(RUSAGE_SELF, &used) == -1) {
674 perror("getrusage(RUSAGE_SELF)");
680 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
681 perror("getrlimit(RLIMIT_MEMLOCK)");
685 printf(", using %ld / %ld MB lockable memory (%.1f%%)",
686 long(used.ru_maxrss / 1024),
687 long(limit.rlim_cur / 1048576),
688 float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
690 printf(", using %ld MB memory (not locked)",
691 long(used.ru_maxrss / 1024));
698 if (should_cut.exchange(false)) { // Test and clear.
699 video_encoder->do_cut(frame);
703 // Reset every 100 frames, so that local variations in frame times
704 // (especially for the first few frames, when the shaders are
705 // compiled etc.) don't make it hard to measure for the entire
706 // remaining duration of the program.
707 if (frame == 10000) {
715 resource_pool->clean_context();
718 void Mixer::get_one_frame_from_each_card(unsigned master_card_index, CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS], bool has_new_frame[MAX_VIDEO_CARDS], int num_samples[MAX_VIDEO_CARDS])
721 // The first card is the master timer, so wait for it to have a new frame.
722 // TODO: Add a timeout.
723 unique_lock<mutex> lock(bmusb_mutex);
724 cards[master_card_index].new_frames_changed.wait(lock, [this, master_card_index]{ return !cards[master_card_index].new_frames.empty() || cards[master_card_index].capture->get_disconnected(); });
726 if (cards[master_card_index].new_frames.empty()) {
727 // We were woken up, but not due to a new frame. Deal with it
729 assert(cards[master_card_index].capture->get_disconnected());
730 handle_hotplugged_cards();
734 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
735 CaptureCard *card = &cards[card_index];
736 if (card->new_frames.empty()) {
737 assert(card_index != master_card_index);
738 card->queue_length_policy.update_policy(-1);
741 new_frames[card_index] = move(card->new_frames.front());
742 has_new_frame[card_index] = true;
743 card->new_frames.pop();
744 card->new_frames_changed.notify_all();
746 int num_samples_times_timebase = OUTPUT_FREQUENCY * new_frames[card_index].length + card->fractional_samples;
747 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
748 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
749 assert(num_samples[card_index] >= 0);
751 if (card_index == master_card_index) {
752 // We don't use the queue length policy for the master card,
753 // but we will if it stops being the master. Thus, clear out
754 // the policy in case we switch in the future.
755 card->queue_length_policy.reset(card_index);
757 // If we have excess frames compared to the policy for this card,
758 // drop frames from the head.
759 card->queue_length_policy.update_policy(card->new_frames.size());
760 while (card->new_frames.size() > card->queue_length_policy.get_safe_queue_length()) {
761 card->new_frames.pop();
767 void Mixer::handle_hotplugged_cards()
769 // Check for cards that have been disconnected since last frame.
770 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
771 CaptureCard *card = &cards[card_index];
772 if (card->capture->get_disconnected()) {
773 fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
774 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
775 configure_card(card_index, capture, /*is_fake_capture=*/true);
776 card->queue_length_policy.reset(card_index);
777 card->capture->start_bm_capture();
781 // Check for cards that have been connected since last frame.
782 vector<libusb_device *> hotplugged_cards_copy;
784 lock_guard<mutex> lock(hotplug_mutex);
785 swap(hotplugged_cards, hotplugged_cards_copy);
787 for (libusb_device *new_dev : hotplugged_cards_copy) {
788 // Look for a fake capture card where we can stick this in.
789 int free_card_index = -1;
790 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
791 if (cards[card_index].is_fake_capture) {
792 free_card_index = int(card_index);
797 if (free_card_index == -1) {
798 fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
799 libusb_unref_device(new_dev);
801 // BMUSBCapture takes ownership.
802 fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
803 CaptureCard *card = &cards[free_card_index];
804 BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
805 configure_card(free_card_index, capture, /*is_fake_capture=*/false);
806 card->queue_length_policy.reset(free_card_index);
807 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
808 capture->start_bm_capture();
814 void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame)
816 // Resample the audio as needed, including from previously dropped frames.
817 assert(num_cards > 0);
818 for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
819 const bool dropped_frame = (frame_num != dropped_frames);
821 // Signal to the audio thread to process this frame.
822 // Note that if the frame is a dropped frame, we signal that
823 // we don't want to use this frame as base for adjusting
824 // the resampler rate. The reason for this is that the timing
825 // of these frames is often way too late; they typically don't
826 // “arrive” before we synthesize them. Thus, we could end up
827 // in a situation where we have inserted e.g. five audio frames
828 // into the queue before we then start pulling five of them
829 // back out. This makes ResamplingQueue overestimate the delay,
830 // causing undue resampler changes. (We _do_ use the last,
831 // non-dropped frame; perhaps we should just discard that as well,
832 // since dropped frames are expected to be rare, and it might be
833 // better to just wait until we have a slightly more normal situation).
834 unique_lock<mutex> lock(audio_mutex);
835 bool adjust_rate = !dropped_frame;
836 audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate});
837 audio_task_queue_changed.notify_one();
840 // For dropped frames, increase the pts. Note that if the format changed
841 // in the meantime, we have no way of detecting that; we just have to
842 // assume the frame length is always the same.
843 pts_int += length_per_frame;
848 void Mixer::render_one_frame(int64_t duration)
850 // Get the main chain from the theme, and set its state immediately.
851 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
852 EffectChain *chain = theme_main_chain.chain;
853 theme_main_chain.setup_chain();
854 //theme_main_chain.chain->enable_phase_timing(true);
856 GLuint y_tex, cbcr_tex;
857 bool got_frame = video_encoder->begin_frame(&y_tex, &cbcr_tex);
860 // Render main chain.
861 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, global_flags.width, global_flags.height);
862 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, global_flags.width, global_flags.height); // Saves texture bandwidth, although dithering gets messed up.
863 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
865 chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
866 resource_pool->release_fbo(fbo);
868 subsample_chroma(cbcr_full_tex, cbcr_tex);
869 resource_pool->release_2d_texture(cbcr_full_tex);
871 // Set the right state for rgba_tex.
872 glBindFramebuffer(GL_FRAMEBUFFER, 0);
873 glBindTexture(GL_TEXTURE_2D, rgba_tex);
874 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
875 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
876 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
878 const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE); // Corresponds to the delay in ResamplingQueue.
879 RefCountedGLsync fence = video_encoder->end_frame(pts_int + av_delay, duration, theme_main_chain.input_frames);
881 // The live frame just shows the RGBA texture we just rendered.
882 // It owns rgba_tex now.
883 DisplayFrame live_frame;
884 live_frame.chain = display_chain.get();
885 live_frame.setup_chain = [this, rgba_tex]{
886 display_input->set_texture_num(rgba_tex);
888 live_frame.ready_fence = fence;
889 live_frame.input_frames = {};
890 live_frame.temp_textures = { rgba_tex };
891 output_channel[OUTPUT_LIVE].output_frame(live_frame);
893 // Set up preview and any additional channels.
894 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
895 DisplayFrame display_frame;
896 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state); // FIXME: dimensions
897 display_frame.chain = chain.chain;
898 display_frame.setup_chain = chain.setup_chain;
899 display_frame.ready_fence = fence;
900 display_frame.input_frames = chain.input_frames;
901 display_frame.temp_textures = {};
902 output_channel[i].output_frame(display_frame);
906 void Mixer::audio_thread_func()
908 while (!should_quit) {
912 unique_lock<mutex> lock(audio_mutex);
913 audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
917 task = audio_task_queue.front();
918 audio_task_queue.pop();
921 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
922 task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
923 vector<float> samples_out = audio_mixer.get_output(
924 double(task.pts_int) / TIMEBASE,
926 rate_adjustment_policy);
928 // Send the samples to the sound card, then add them to the output.
930 alsa->write(samples_out);
932 video_encoder->add_audio(task.pts_int, move(samples_out));
936 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
939 glGenVertexArrays(1, &vao);
942 glBindVertexArray(vao);
946 GLuint fbo = resource_pool->create_fbo(dst_tex);
947 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
948 glViewport(0, 0, global_flags.width/2, global_flags.height/2);
951 glUseProgram(cbcr_program_num);
954 glActiveTexture(GL_TEXTURE0);
956 glBindTexture(GL_TEXTURE_2D, src_tex);
958 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
960 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
962 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
965 float chroma_offset_0[] = { -1.0f / global_flags.width, 0.0f };
966 float chroma_offset_1[] = { -0.0f / global_flags.width, 0.0f };
967 set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
968 set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_1", chroma_offset_1);
970 glUniform1i(cbcr_texture_sampler_uniform, 0);
972 glBindBuffer(GL_ARRAY_BUFFER, cbcr_vbo);
975 for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
976 glEnableVertexAttribArray(attr_index);
978 glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
982 glDrawArrays(GL_TRIANGLES, 0, 3);
985 for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
986 glDisableVertexAttribArray(attr_index);
992 glBindFramebuffer(GL_FRAMEBUFFER, 0);
995 resource_pool->release_fbo(fbo);
996 glDeleteVertexArrays(1, &vao);
999 void Mixer::release_display_frame(DisplayFrame *frame)
1001 for (GLuint texnum : frame->temp_textures) {
1002 resource_pool->release_2d_texture(texnum);
1004 frame->temp_textures.clear();
1005 frame->ready_fence.reset();
1006 frame->input_frames.clear();
1011 mixer_thread = thread(&Mixer::thread_func, this);
1012 audio_thread = thread(&Mixer::audio_thread_func, this);
1018 audio_task_queue_changed.notify_one();
1019 mixer_thread.join();
1020 audio_thread.join();
1023 void Mixer::transition_clicked(int transition_num)
1025 theme->transition_clicked(transition_num, pts());
1028 void Mixer::channel_clicked(int preview_num)
1030 theme->channel_clicked(preview_num);
1033 void Mixer::start_mode_scanning(unsigned card_index)
1035 assert(card_index < num_cards);
1036 if (is_mode_scanning[card_index]) {
1039 is_mode_scanning[card_index] = true;
1040 mode_scanlist[card_index].clear();
1041 for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1042 mode_scanlist[card_index].push_back(mode.first);
1044 assert(!mode_scanlist[card_index].empty());
1045 mode_scanlist_index[card_index] = 0;
1046 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1047 last_mode_scan_change[card_index] = steady_clock::now();
1050 Mixer::OutputChannel::~OutputChannel()
1052 if (has_current_frame) {
1053 parent->release_display_frame(¤t_frame);
1055 if (has_ready_frame) {
1056 parent->release_display_frame(&ready_frame);
1060 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1062 // Store this frame for display. Remove the ready frame if any
1063 // (it was seemingly never used).
1065 unique_lock<mutex> lock(frame_mutex);
1066 if (has_ready_frame) {
1067 parent->release_display_frame(&ready_frame);
1069 ready_frame = frame;
1070 has_ready_frame = true;
1073 if (new_frame_ready_callback) {
1074 new_frame_ready_callback();
1077 // Reduce the number of callbacks by filtering duplicates. The reason
1078 // why we bother doing this is that Qt seemingly can get into a state
1079 // where its builds up an essentially unbounded queue of signals,
1080 // consuming more and more memory, and there's no good way of collapsing
1081 // user-defined signals or limiting the length of the queue.
1082 if (transition_names_updated_callback) {
1083 vector<string> transition_names = global_mixer->get_transition_names();
1084 bool changed = false;
1085 if (transition_names.size() != last_transition_names.size()) {
1088 for (unsigned i = 0; i < transition_names.size(); ++i) {
1089 if (transition_names[i] != last_transition_names[i]) {
1096 transition_names_updated_callback(transition_names);
1097 last_transition_names = transition_names;
1100 if (name_updated_callback) {
1101 string name = global_mixer->get_channel_name(channel);
1102 if (name != last_name) {
1103 name_updated_callback(name);
1107 if (color_updated_callback) {
1108 string color = global_mixer->get_channel_color(channel);
1109 if (color != last_color) {
1110 color_updated_callback(color);
1116 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1118 unique_lock<mutex> lock(frame_mutex);
1119 if (!has_current_frame && !has_ready_frame) {
1123 if (has_current_frame && has_ready_frame) {
1124 // We have a new ready frame. Toss the current one.
1125 parent->release_display_frame(¤t_frame);
1126 has_current_frame = false;
1128 if (has_ready_frame) {
1129 assert(!has_current_frame);
1130 current_frame = ready_frame;
1131 ready_frame.ready_fence.reset(); // Drop the refcount.
1132 ready_frame.input_frames.clear(); // Drop the refcounts.
1133 has_current_frame = true;
1134 has_ready_frame = false;
1137 *frame = current_frame;
1141 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1143 new_frame_ready_callback = callback;
1146 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1148 transition_names_updated_callback = callback;
1151 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1153 name_updated_callback = callback;
1156 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1158 color_updated_callback = callback;
1161 mutex RefCountedGLsync::fence_lock;